Deep learning-based heart age prediction bias correction method

Abstract

A deep learning-based method for correcting cardiac age prediction bias comprises the following steps: Step 1: Acquire 12-lead ECG data and actual age information; filter the ECG data to construct healthy and disease sample sets; Step 2: Preprocess the acquired ECG data; Step 3: Construct a cardiac age prediction model and train and validate it; Step 4: Perform residual correction on the cardiac age prediction data output by the cardiac age prediction model to obtain a correction coefficient lookup table; and obtain the corrected predicted cardiac age according to the correction coefficient lookup table, thus obtaining the corrected age difference value. This invention first corrects the AG (age difference) for each age group through residual correction, eliminating the problems of overestimation in younger ages and underestimation in older ages. It has strong generalization ability, and can stably correct biases in all age groups, especially in extreme age ranges, where the residual correction effect is superior to traditional linear correction methods.

Description

Technical Field

[0001] This invention relates to the field of medical and health care and artificial intelligence interdisciplinary technology, and more specifically, to a method for correcting the bias in heart age prediction based on deep learning. Background Technology

[0002] Electrocardiography (ECG), as a non-invasive method for recording cardiac electrical activity, effectively reflects cardiac function and individual health status. The Framingham Heart Study in 2008 introduced the concept of cardiac age, a crucial indicator for assessing cardiovascular health. The difference between cardiac age and chronological age (AgeGap, AG) reflects the degree of cardiac aging, and its clinical value far exceeds the predictive value of cardiac age alone. A cardiac age greater than chronological age, with a positive age difference, suggests accelerated cardiac aging and is a significant warning sign of increased cardiovascular disease incidence and mortality. Conversely, a cardiac age less than or equal to chronological age, with a zero or negative age difference, indicates a cardiac state that matches or is younger than chronological age, suggesting lower health risk. Therefore, age difference, as a core biomarker, plays an irreplaceable role in cardiovascular risk stratification, early disease screening, and evaluation of clinical intervention effectiveness.

[0003] In recent years, deep neural networks have been widely used to extract features from raw ECG data and directly predict cardiac age. One type is based on statistical models using traditional risk factors, which establish predictive equations by integrating clinical parameters such as heart rate and blood pressure. The other type is based on end-to-end models using deep neural networks, which directly map ECG time-series signals to predicted cardiac age values. This can lead to underestimating cardiac age in older individuals and overestimating cardiac age in younger individuals, resulting in misjudgments of disease risk. While these technologies have improved prediction accuracy, none of them specifically correct for the systematic bias in age differences. They treat cardiac age prediction as the ultimate goal, completely ignoring the reliability requirements of age differences as a core clinical indicator.

[0004] Existing methods for correcting cardiac age differences have several shortcomings. First, current research focuses primarily on improving the accuracy of cardiac age predictions, with most attempts using global linear regression, which only corrects overall trend deviations and fails to capture the nonlinear relationship between age and deviations. Second, neuroscience-based brain age correction methods suffer from poor direct transfer effects and cross-domain adaptability due to fundamental differences in the physiological mechanisms and characteristic dimensions of EEG and ECG signals. Third, training data often includes both healthy and diseased samples, and the pathological features associated with diseases can interfere with the model's learning of normal cardiac aging trajectories, causing the predicted cardiac age to deviate from the physiological truth, thus rendering the age difference meaningless in reflecting the core rate of physiological aging. Fourth, the lack of age grouping for local bias adjustment results in weak clinical discriminative power.

[0005] Therefore, the existing technology has problems and needs further improvement and development. Summary of the Invention

[0006] (I) Purpose of the invention: In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a method for correcting the deviation of heart age prediction based on deep learning, which aims to correct the deviation of age difference, improve the clinical application value of heart age as a biomarker of cardiovascular health, and solve the technical problem that the existing technology cannot meet the clinical requirements for accurate assessment of age difference.

[0007] (II) Technical Solution: To solve the above-mentioned technical problems, this technical solution provides a deep learning-based method for correcting biases in cardiac age prediction, including:

[0008] Step 1: Obtain 12-lead ECG data and actual age information, filter the ECG data, and construct a healthy sample set and a disease sample set;

[0009] Step 2: Preprocess the acquired ECG data;

[0010] Step 3: Build a cardiac age prediction model and train and validate it;

[0011] Step 4: Perform residual correction on the heart age prediction data output by the heart age prediction model to obtain a correction coefficient lookup table. Obtain the corrected predicted heart age according to the correction coefficient lookup table, and thus obtain the corrected age difference value.

[0012] Preferably, step 1 is as follows:

[0013] Step 101: Obtain multi-source data from multiple databases;

[0014] Step 102: Select normal samples according to clinical standards and remove noise, artifacts and missing signal samples from the 12-lead ECG data;

[0015] Step 103: Perform age balancing on the normal samples to construct healthy sample sets and disease sample sets respectively.

[0016] Preferably, samples of subjects aged 16 to 85 years are selected for age balancing, and samples are randomly drawn from each age group until the total sample limit of 2000 for that age group is reached; when the sample size of a certain age group is insufficient, all samples are included.

[0017] The healthy sample set was divided into training set, validation set and test set in an 8:1:1 ratio, and the distribution ratio of each age group in the training set, validation set and test set remained consistent.

[0018] The disease sample sets include LBBB sample sets, RBBB sample sets, HF sample sets, and LVH sample sets.

[0019] Preferably, when the application layer receives ECG data, it calls the underlying artifact detection request entry module. The artifact detection request entry module parses the ECG data, separates the ECG signal and the custom configuration parameters, and sends the ECG signal to the dual-dimensional merging processing module and the custom configuration parameters to the artifact priority configuration module.

[0020] The dual-dimensional merging processing module shares judgment parameters for similar artifact instances within a continuous time interval of the same lead, only recording the start and end times; for artifact instances in different leads but with matching features, it uses unified detection logic to generate detection clusters and passes them to the hierarchical storage management module; the hierarchical storage management module stores core lightweight instances in a high-speed cache, and merges detection clusters and custom threshold instances in secondary storage; during detection, it first loads core instances from the high-speed cache, and if no match is found, it loads detection clusters / low-frequency instances from the secondary storage, and sends the matched artifact instances / detection clusters to the hierarchical voxelization detection module;

[0021] The artifact priority configuration module combines clinical default rules with user-defined configurations to generate a priority strategy for the current detection and sends it to the graded voxelization detection module. The graded voxelization detection module performs high-precision detection on high-priority artifacts and voxelization detection on low-priority artifacts based on the received artifact instances / detection clusters and priority strategy. At the same time, it adjusts the detection precision according to system resource usage and outputs the artifact detection results to the application layer.

[0022] Preferably, step 2 includes:

[0023] Step 201: Resample the 12-lead ECG data to 500Hz and crop it to a fixed duration of 10 seconds;

[0024] Step 202: The amplitude of the 12-lead ECG data signal is normalized and arranged in the order of I, II, III, aVR, aVL, aVF, V1-V6 to form the model input data.

[0025] Preferably, the 12-lead data signals are arranged in a fixed order of I, II, III, aVR, aVL, aVF, V1-V6, and the channels are reorganized to form a (12, 5000) signal matrix.

[0026] Preferably, step 3 includes:

[0027] Step 301: Construct a cardiac age prediction model;

[0028] Step 302: Train the heart age prediction model;

[0029] Step 303: Validate the cardiac age prediction model.

[0030] Preferably, the heart age prediction model includes an initial convolutional block, four stacked residual blocks, a global average pooling layer, and a fully connected output layer;

[0031] The heart age model is trained using actual age as input and heart age as output, with weighted mean square error as loss function, Adam optimizer, and ReduceLROnPlateau learning rate decay strategy.

[0032] Monitor changes in the loss value on the validation set and employ an early stopping strategy to prevent overfitting.

[0033] Preferably, step 4 includes:

[0034] Step 401: Group the age range by age;

[0035] Step 402: Calculate the deviation coefficient for each age group;

[0036] Step 403: Apply the bias correction coefficient to correct the bias of the sample to be corrected, and calculate the corrected age difference.

[0037] Preferably, the age range [age] min ,age max [Divided sequentially into K consecutive age groups G] j The width of each age group is a fixed value w, and the width of the j-th age group G is... j =[age min +j×w,age min +(j+1)×w],j=0,1,…,K-1,w is the span of each age group,w=3~10.

[0038] Preferably, for each sample i in the validation set, the true age is y. A,i The predicted age of the cardiac age prediction model is y. A,pred,i The samples were sorted according to their corresponding real ages y. A,i Classified into the corresponding age group G j In the middle, for each age group G j Calculate all N contained within it j The average predicted residual of each sample is the bias correction factor b for that age group. j : The correction coefficient lookup table {G} is obtained. j :b j};

[0039] For the sample m to be corrected, its actual age is y. B,mThe original predicted heart age of the heart age prediction model is y. pred,m Determine its age group G. j Find the G value for this age group from the correction coefficient lookup table. j The corresponding deviation coefficient b j The original predicted cardiac age y of the sample m to be corrected pred,m Subtract the corresponding deviation correction factor b j The corrected heart age y was obtained. corr,m y corr,m =y pred,m -b j；

[0040] The corrected age difference (AG) is calculated based on the corrected cardiac age and the original predicted cardiac age. corr AG corr =y corr,m -y B,m。

[0041] (III) Beneficial Effects: This invention provides a deep learning-based method for correcting cardiac age prediction bias. First, it corrects the AG (Advanced Age Scale) for each age group through a bias correction coefficient, eliminating the problem of overestimation in younger individuals and underestimation in older individuals, thus achieving a physiological state that matches cardiac age with actual age and freeing the AG index from the interference of age-related errors. Second, it significantly improves the T-scores between the disease group and the healthy group, making the distinction between diseased and healthy individuals more significant, enabling the AG to more accurately capture pathological cardiac aging caused by disease. Third, the risk discrimination of the corrected AG is significantly better than that before correction, providing strong support for precise clinical risk stratification and targeted intervention for high-risk individuals, making it a superior survival prediction biomarker. Finally, in extreme age ranges, the residual correction effect is better than the traditional linear correction method, effectively solving the prediction reliability problem of the model in the edge region of data distribution, laying the foundation for cross-population and cross-institutional clinical applications. Attached Figure Description

[0042] Figure 1 This is a flowchart of the steps of the deep learning-based cardiac age prediction bias correction method of the present invention;

[0043] Figure 2 This is a schematic diagram of the cardiac age prediction model.

[0044] Figure 3 This is a graph showing the loss function for validating the heart age prediction model;

[0045] Figure 4 This is a flowchart of the residual correction process;

[0046] Figure 5 This is a graph showing the average age difference across different age groups before residual correction.

[0047] Figure 6 This is a graph showing the average age difference across different age groups after residual correction.

[0048] Figure 7 This is a heatmap of the heart age prediction model on the training set of a healthy sample set;

[0049] Figure 8 This is a heatmap of the cardiac age prediction model on the validation set of a healthy sample set;

[0050] Figure 9 This is a heatmap of the heart age prediction model on the test set of a healthy sample set;

[0051] Figure 10 This is a risk prediction chart for age groups before and after residual correction;

[0052] Figure 11 This is a survival rate curve under the optimal threshold before residual correction;

[0053] Figure 12 This is a survival rate curve at the optimal threshold after residual correction;

[0054] Figure 13 This is a comparison chart of different correction methods and the results before and after correction. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to preferred embodiments. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0056] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that these drawings are for illustrative purposes only and are not drawn to scale, and should not be construed as limiting the actual scope of protection of the present invention.

[0057] Deep learning-based methods for correcting biases in cardiac age prediction, such as Figure 1 As shown, it specifically includes:

[0058] Step 1: Obtain 12-lead ECG data and actual age information, filter the ECG data, and construct a healthy sample set and a disease sample set.

[0059] Step 2: Preprocess the acquired ECG data.

[0060] Step 3: Build a heart age prediction model and train and validate it.

[0061] Step 4: Perform residual correction on the cardiac age prediction data output by the cardiac age prediction model to obtain the corrected age difference value.

[0062] More specifically, step 1 is as follows:

[0063] Step 101: Obtain multi-source data from multiple databases.

[0064] 12-lead ECG data and the subjects' actual age information were collected from four public ECG databases: CODE-15%, MIMIC-IV, PTB-XL, and UK Biobank.

[0065] By using Python libraries such as wfdb (reading WFDB format), scipy.io (reading MAT format), and pydicom (reading DICOM format), key information such as raw signal data, sampling frequency, and recording duration of 12-lead ECG data can be extracted.

[0066] Step 102: Select normal samples according to clinical standards and remove samples with noise, artifacts and missing signals from the 12-lead ECG data.

[0067] The normal samples must meet the basic requirements for ECG data quality, have bands within the normal range, and be verified by clinical diagnostic annotations.

[0068] The basic requirements for ECG data quality are signal integrity, a signal-to-noise ratio higher than the signal strength, and being within the heart rate range. Signal integrity means that there is no missing data in the 12-lead ECG, the effective recording time of each lead ECG data is not less than the preset analysis time, which is preferably 10 seconds in this invention, and there is no flat baseline or no signal segment caused by ECG data loss from any lead.

[0069] The signal-to-noise ratio (SNR) is used to quantify and assess the noise level. The requirement is that the SNR is ≥30dB, and low-quality signals caused by electromyography interference, power frequency interference, and motion artifacts should be excluded.

[0070] The normal sinus rhythm range is 60–100 beats per minute (bpm). The heart rate is calculated by the R wave peak interval in the ECG data, and samples with bradycardia, tachycardia, and irregular heart rhythm are excluded.

[0071] The normal range for wave amplitude includes P wave amplitude <0.25mV and duration <0.12 seconds; QRS amplitude within the normal range, for example, R wave 0.5-2.5mV in lead I, S wave <3.0mV in lead V1, and R wave <2.5mV in lead V5, with no pathological Q waves; no ST segment elevation or depression, and no flattened, inverted, or peaked T waves; QT interval, after heart rate correction, is <0.45 seconds for men and <0.47 seconds for women.

[0072] The clinical diagnostic labeling verification refers to combining the clinical information of subjects in the database to exclude samples with a history of cardiovascular disease, hypertension, diabetes, electrolyte imbalance, and those taking drugs that affect cardiac electrical activity, to ensure that there is no pathological interference.

[0073] When the application layer data receiving module receives ECG data, the underlying system's track segmentation module breaks down the received ECG data into multiple time tracks for filtering. All information of the ECG data is presented intuitively on the same interface, allowing users to view artifacts, signal quality, and the quality of pre-processed ECG data without switching tools, thus improving efficiency.

[0074] The multi-layer time track comprises four types of tracks: raw signal track, artifact marker track, signal quality track, and preprocessing parameter track. The raw signal track refers to the 12-lead ECG data as independent tracks, arranged in the order of I, II, III, aVR, aVL, aVF, and V1-V6, displaying the original time-series signal. The artifact marker track uses different colors to indicate artifact types and start / end times; for example, in this invention, red represents motion artifacts, blue represents electrode artifacts, and green represents electromyography artifacts. The signal quality track uses different colors to indicate the quality of the ECG data signal; in this invention, gray blocks mark lead missing regions with no effective signal for more than 2 consecutive seconds, orange bars mark samples shorter than 10 seconds, and pink blocks mark qualified signal segments with a signal-to-noise ratio ≥30dB. The preprocessing parameter track records key preprocessing configurations such as resampling frequency, amplitude normalization parameters, and filtering parameters, and correlates them in real time with signal changes within the corresponding time interval. When adjusting preprocessing parameters, the effect of changes in the signal quality track is previewed in real time.

[0075] The multi-layered time track uses a multi-dimensional time axis as the main line to realize full-process operation on a single interface, and supports time window scaling, customized guide track layout, and real-time preview of strategy parameters.

[0076] The timeline automatically identifies missing lead regions, allowing users to manually add interpolation or mark removal. It features quick keyframe jumps and timeline snapshot comparisons, with processing results synchronized to the original signal track in real time. By selecting artifact-marked regions on the timeline, targeted filtering is applied to the ECG data within those regions, or invalid samples with artifact percentages exceeding 30% are directly removed. When parameters such as resampling frequency and filtering thresholds are modified, the signal quality track updates the processing effect in real time, avoiding inefficient operations caused by repeated testing.

[0077] Noise removal employs a wavelet transform-based noise identification method, which decomposes ECG data into multiple scales to extract high-frequency and low-frequency noise components. When the noise energy percentage is greater than 10%, the sample is identified as noise and removed.

[0078] The artifacts include motion artifacts, electrode artifacts, and electromyography (EMG) artifacts. Motion artifacts are identified by calculating the baseline drift velocity; if the drift duration exceeds 1 second, or if there are sudden signal jumps, it is considered a motion artifact. Electrode artifacts refer to signal interruptions or high-frequency spikes caused by poor electrode contact. EMG artifacts are identified when the high-frequency component energy percentage in the signal exceeds 15%, and the R-wave recognition accuracy is less than 95%.

[0079] When the application layer receives ECG data, it calls the underlying artifact detection request entry module. This module parses the ECG data from the application layer, separating the ECG signal from the custom configuration parameters. The ECG signal is then sent to the dual-dimensional merging processing module, and the custom configuration parameters are sent to the artifact priority configuration module. For similar artifact instances within the same lead's continuous time interval, the dual-dimensional merging processing module shares the judgment parameters and only records the start and end times. For artifact instances in different leads but with matching features, it uses a unified detection logic to avoid redundant parameter loading, generating detection clusters and passing them to the hierarchical storage management module. The hierarchical storage management module receives the detection clusters and unmerged artifact instances from the dual-dimensional merging processing module. It stores the core lightweight instances in a cache and stores the merged detection clusters and custom threshold instances in secondary storage. During detection, the core instances are first loaded from the cache; if no match is found, the detection clusters / low-frequency instances from the secondary storage are loaded to reduce invalid memory usage. Matched artifact instances / detection clusters are then sent to the hierarchical voxelization detection module.

[0080] The artifact priority configuration module receives the custom configuration parameters from the artifact detection request entry module, combines the clinical default rules with the user-defined configuration, generates the priority strategy for the current detection, and sends the detection priority to the graded voxelization detection module.

[0081] The hierarchical voxelization detection module performs high-precision detection on high-priority artifacts and simplified voxelization detection on low-priority artifacts based on the received artifact instances / detection clusters and priority strategy. It also adjusts the detection precision according to system resource usage and outputs the artifact detection results to the application layer.

[0082] When detecting artifacts, the occupancy rate of inner layers should be reduced, including lightweight instantiation merging strategies, hierarchical voxelization processing strategies, and dynamic calculation strategies for artifact contour replacement.

[0083] Specifically, the lightweight instantiation merging strategy refers to the dual-dimensional merging processing module merging detection rules for similar artifacts into lightweight instances, each containing only the necessary detection parameters; adjacent instances of the same type can be merged into larger detection clusters, reducing memory usage and redundant computation overhead. The dual-dimensional merging processing module outputs the generated detection clusters to the hierarchical storage management module for hierarchical storage.

[0084] For motion artifacts, electrode artifacts, and electromyography artifacts, the dual-dimensional merging module encapsulates the detection rules for each type of artifact into independent lightweight instances, retaining only the necessary parameters. Each instance contains only the artifact type identifier, core judgment parameters, and output result template, without any additional redundant data. The parameter data size of each lightweight instance is strictly controlled within the range of 128 bytes to 1KB to avoid low merging efficiency due to excessively small granularity or rigid detection due to excessively large granularity.

[0085] The dual-dimensional merging module employs a two-dimensional merging logic based on spatial adjacency and feature similarity to integrate multiple lightweight instances into a larger detection cluster, reducing redundant computation overhead. For similar artifact instances within the same lead and continuous time interval, temporal adjacency is used to merge them into a single detection cluster, sharing core decision parameters and only recording the start and end times of different segments. For instances in different leads but with similar artifact features, feature similarity is used to merge them into a single detection cluster, uniformly executing the detection logic and avoiding repeated parameter loading for each lead.

[0086] Each detection cluster contains a cluster identifier, shared parameters, member instance index, and merging rules. It supports dynamically adding or removing instances, and the merged detection cluster does not change the core judgment logic of a single instance.

[0087] The dual-dimensional merging processing module outputs the core lightweight instances of the three types of artifacts to the hierarchical storage management module for high-speed caching. During invocation, these instances are read directly without disk input or / or output. The dual-dimensional merging processing module also outputs the merged detection clusters and instances with custom thresholds to the hierarchical storage management module for secondary storage. During invocation, these instances are quickly located and loaded based on cluster and instance indices. During artifact detection, core instances in the high-speed cache are loaded first. If no matching artifact is found, the detection clusters and low-frequency instances in the secondary storage are then loaded, reducing unnecessary loading of inner-layer instances.

[0088] The hierarchical voxelization processing strategy refers to the artifact priority configuration module employing high-precision detection logic for high-priority artifacts and simplified voxelization detection logic for low-priority artifacts, locking only key feature contours. While maintaining artifact recognition accuracy, the artifact priority configuration module reduces the rasterization computation overhead of low-priority artifacts by at least 30%. The voxelization switch can be configured via system parameters, and the engine processes the simplified logic using a double-sided single-layer model to improve speed. Users can also customize artifact priorities according to clinical scenarios, flexibly adapting to different detection needs.

[0089] The artifact priority configuration module, considering clinical needs and detection difficulty, classifies motion artifacts, electrode artifacts, and electromyography artifacts into high and low priorities. High-priority artifacts include electrode artifacts; signal interruptions / high-frequency spikes can easily lead to misinterpretation of key waveforms, affecting the calculation of parameters such as heart rate, and require high-precision detection. The high-priority artifacts employ a complete high-precision detection process, analyzing signal waveforms frame by frame, calculating spike amplitudes, and verifying signal interruption duration to ensure a detection accuracy of ≥98%.

[0090] Low-priority artifacts, including minor electromyography (EMG) artifacts and slow-motion artifacts, are identified using a simplified voxelization logic, including feature contour extraction, voxelization modeling, and fast matching. Feature contour extraction ignores subtle fluctuations and only extracts the core contour features of the artifacts. Voxelization modeling maps the extracted contour features to a low-dimensional voxel mesh, using the activation or deactivation state of voxels to characterize the presence of artifacts, replacing frame-by-frame parameter calculation. Fast matching pre-stores voxelization templates for low-priority artifacts; during detection, the similarity between the voxelization result of the input signal and the template is directly compared. A similarity ≥ 85% is considered a corresponding artifact.

[0091] The hierarchical voxelization detection module adjusts the voxel grid size according to the artifact type. For electromyography (EMG) artifacts, which emphasize frequency range characterization, a 16×16×1 grid is used; for motion artifacts, which emphasize temporal trends, an 8×32×1 grid is used. This module monitors inner-layer resource usage in real time. When memory / CPU usage exceeds 70%, it automatically expands the voxelization simplification range for low-priority artifacts, for example, compressing the voxel grid of EMG artifacts to 8×8×1. When usage is below 70%, it can improve the voxelization accuracy of low-priority artifacts, for example, expanding the voxel grid of EMG artifacts to 32×32×1.

[0092] The artifact contour replacement dynamic calculation strategy refers to the dynamic calculation module using a pre-calculated artifact contour rule library instead of real-time dynamic calculation, without deviating from the fixed detection pipeline; the artifact detection rules in the artifact contour rule library are pre-stored in memory, improving matching efficiency; and the rule library can be updated in real time through an interactive optimization unit to avoid contour loss.

[0093] Specifically, the rule base management module breaks down the detection logic for motion artifacts, electrode artifacts, and electromyography artifacts into main rules and branch rules. The main rules refer to the core judgment conditions for each type of artifact and are fixedly stored in the core layer of the rule base; the branch rules refer to the subdivided scene variations of the artifacts. For each type of artifact, the rule base management module pre-calculates threshold boundaries and feature templates for the main / branch rules and stores them as structured data. Main rules are stored in a high-speed cache, and branch rules are stored in secondary storage. During invocation, they are loaded in the order of matching the main rule to the appropriate branch rule, reducing the occupancy of inner layers. The rule base management module uniformly outputs condition parameters, judgment logic, and standardized output results in a standardized format, ensuring that the rule base matching module does not need dynamic parsing when performing detection rule matching, thus improving matching speed.

[0094] When the application layer system receives sample data, it initiates the artifact detection request entry module of the underlying system. This module triggers artifact detection based on the sample data. First, it extracts key features from the ECG data and outputs them to the rule base matching module. The rule base matching module matches the corresponding main rule from the rule base and then adapts branch rules based on feature details. The matching result directly outputs the artifact type and start / end time to the application layer system. If the rule base matching module does not find a matching rule, it initiates lightweight dynamic computation to supplement the detection, avoiding missed detections caused by relying entirely on predefined rules.

[0095] Based on memory or CPU usage, the hierarchical voxelization detection module automatically adjusts the instance merging granularity and hierarchical processing precision for artifact detection. When memory usage exceeds 80%, the hierarchical voxelization detection module expands the instance merging granularity and reduces the detection precision of low-priority artifacts; conversely, it reduces the granularity and increases the precision.

[0096] The hierarchical storage management module caches historical artifact detection results for the same lead. When an artifact with the same characteristics is detected again, the cached result is reused directly. The cache validity period is configurable, and the default value in this invention is 30 seconds.

[0097] If any lead has no valid signal for more than 2 consecutive seconds, or if the length of the 12-lead ECG data is less than 10 seconds, it is determined to be a sample with missing lead signal. If key waveforms such as the R wave and P wave in the ECG data cannot be effectively identified, resulting in the inability to calculate basic parameters such as heart rate and PR interval, it is determined to be a sample with missing key features and is removed.

[0098] Step 103: Perform age balancing on the normal samples to construct healthy sample sets and disease sample sets respectively.

[0099] The underlying system's interactive optimization unit also employs a dynamic interactive adjustment strategy to improve the quality of the sample set, including fine-grained time window verification and sample screening visualization. The fine-grained time window verification supports scaling any lead data from seconds to milliseconds, accurately identifying suspected noise areas. It also provides a keyframe jump function, enabling rapid location of all artifact markers, signal loss areas, and low-quality signal segments, improving the efficiency of manual review. The sample screening visualization supports splitting the timeline view by age group, with each age group independently displaying the signal quality distribution of its corresponding samples, facilitating age balancing processing.

[0100] A sample of subjects aged 16-85 years was selected for age balancing, and samples were randomly drawn from each age group until the total sample size for that age group reached the upper limit of 2000. If the sample size of a certain age group was insufficient, all samples were included.

[0101] The age balancing process is a key step in constructing a pure physiological aging benchmark. Its core objective is to eliminate the impact of differences in sample size among different age groups on model training, ensure that the model can uniformly learn the physiological aging characteristics of each age group, and avoid prediction bias caused by sample distribution bias.

[0102] Multiple samples were obtained from four databases for separate assessment, constructing a healthy sample set and four disease datasets. Specifically, 193,018 samples were obtained from CODE-15%, 800,035 from MIMIC-IV, 54,962 from UK Biobank, and 21,799 from PTB-XL. These samples were divided into training, validation, and test sets in an 8:1:1 ratio, ensuring consistent age distribution across all age groups to maintain data balance. Simultaneously, four disease datasets were constructed: LBBB (Left Bundle Branch Block), RBBB (Right Bundle Branch Block), HF (Heart Failure), and LVH (Left Ventricular Hypertrophy).

[0103] The specific operations of step 2 include:

[0104] Because ECG data comes from multiple public databases such as CODE-15%, MIMIC-IV, PTB-XL, and UK Biobank, the signal sampling frequency and recording duration of different databases are different, resulting in inconsistent signal dimensions and temporal resolution, making it impossible to directly input them into deep learning models for unified training.

[0105] Step 201: Resample the 12-lead ECG data to 500Hz and crop it to a fixed duration of 10 seconds.

[0106] A linear interpolation resampling method based on Fourier transform is employed to ensure consistent time resolution. After resampling, it is necessary to verify that the core frequency components of the ECG data signal are not lost. Through power spectral density analysis, it is ensured that the signal energy proportion in the 0.5–100 Hz frequency band is ≥95%. It is also necessary to verify that there are no significant changes in key waveform parameters, with the R-wave peak amplitude variation rate ≤5% and the QRS width variation rate ≤3% before and after resampling, to ensure that physiological characteristics are not distorted due to resampling.

[0107] After the resampling verification is completed, the ECG data signal is clipped to a fixed time of 10 seconds to ensure data dimension consistency.

[0108] Step 202: The amplitude of the 12-lead ECG data signal is normalized and arranged in the order of I, II, III, aVR, aVL, aVF, V1-V6 to form the model input data.

[0109] The amplitude units of ECG data were uniformly converted to microvolts (μV). Specifically, all resampled and 10-second cropped ECG data signals were traversed, and the unit labels of the ECG data signals from each database were extracted. Different units were then converted. If the signal unit was millivolts (mV), it was multiplied by 1000 to convert to microvolts (μV). If the unit was not labeled, it was determined by the signal amplitude range. If the original signal amplitude range was between -0.5 and 0.5, it was determined to be in mV units, and a multiplication by 1000 conversion was performed. If the range was between -500 and 500, it was directly determined to be in μV units without conversion. Unit conversion was performed to address the differences in unit labels between different databases, ensuring the comparability of ECG data signal amplitudes.

[0110] Read the lead labeling information from each database, identify the signal channels corresponding to the 12 leads, and establish a mapping table between the original lead names and the target order for some databases where the lead order is disordered, as shown in Table 1, to ensure that each lead accurately matches the target position.

[0111] Table 1. Mapping Table of Original Lead Names and Target Order

[0112]

[0113] Arranged in a fixed order of I, II, III, aVR, aVL, aVF, V1-V6, the 12-lead data signals are reorganized into a (12, 5000) two-dimensional signal matrix, i.e., 12 leads × 5000 data points, to prepare for subsequent model input.

[0114] By standardizing the amplitude to unify the signal magnitude, arranging the 12 leads in a fixed order, and constructing standardized data dimensions, we ensure that the signal format of the input model is uniform and the features are comparable, thus providing a guarantee for the model to stably extract the physiological aging features of the heart.

[0115] Step 3 specifically includes:

[0116] Step 301: Construct a cardiac age prediction model.

[0117] To address the one-dimensional temporal characteristics of 12-lead ECG data signals, a one-dimensional convolutional neural network with residual connections is used to construct a heart age prediction model. This solves the gradient vanishing problem in deep network training and accurately captures subtle temporal features related to heart age in ECG data signals. Figure 2 As shown, the heart age prediction model includes an initial convolutional block, four stacked residual blocks, a global average pooling layer, and a fully connected output layer.

[0118] The initial convolutional block performs preliminary feature extraction, including a one-dimensional convolutional layer, a batch normalization layer, and a dropout layer. The one-dimensional convolutional layer performs preliminary feature extraction on the input 12-lead ECG data signal, preserving temporal resolution, and includes 64 kernels, a kernel size of 3, a stride of 1, padding of 1, and the ReLU activation function. The batch normalization layer accelerates convergence and reduces internal covariate bias, including a dimension of 64, momentum of 0.9, and eps of 1e-5. The dropout layer prevents overfitting and enhances generalization ability, with a dropout rate of 0.2.

[0119] Four residual blocks are stacked for deep feature extraction. Each residual block contains the main path and residual connection paths.

[0120] The main path includes a one-dimensional convolutional layer 1, a batch normalization layer, a dropout layer, a one-dimensional convolutional layer 2, and another batch normalization layer. The one-dimensional convolutional layer 1 has 64 kernels, a kernel size of 3, a stride of 1, padding of 1, and a ReLU activation function. The dropout layer has a dropout rate of 0.2. The one-dimensional convolutional layer 2 has 64 kernels, a kernel size of 3, a stride of 1, padding of 1, and no activation function.

[0121] The residual connection path includes direct identity mapping when the input and output dimensions of the main path are the same; when the dimensions are different, the connection is made after adjusting the dimensions through 1×1 convolution.

[0122] After the main path output and the residual path output are added together, the output is processed by the ReLU activation function. The four residual blocks are stacked to gradually extract the deep temporal features of the 12-lead ECG data signal.

[0123] The global average pooling layer uses the time dimension as its pooling dimension. It calculates the average value of the time-dimensional data for each convolutional channel, compressing the two-dimensional feature tensor of channel × time into a one-dimensional feature vector for each channel. This reduces the number of parameters, avoids overfitting, and preserves the global features of each channel. The global average pooling layer aggregates and reduces the dimensionality of deep convolutional features, adapting to the single-value output requirements of regression tasks.

[0124] The fully connected output layer performs age prediction and includes a fully connected layer 1, a dropout layer, and a fully connected layer 2. Fully connected layer 1 performs a non-linear transformation on the aggregated feature vector to enhance feature representation, and includes 128 neurons and a ReLU activation function. The dropout layer further prevents overfitting with a dropout rate of 0.2. Fully connected layer 2 outputs a single predicted cardiac age, matching the single-value prediction task requirements of actual age and cardiac age, and includes 1 neuron and no activation function.

[0125] Step 302: Train the heart age prediction model.

[0126] The heart age model is trained using actual age as input and cardiac age as output, employing weighted mean squared error as the loss function, the Adam optimizer, and a ReduceLROnPlateau learning rate decay strategy. The training epochs are 70, and the batch size is 32.

[0127] The weight of the weighted mean square error is equal to the reciprocal of the sample proportion of each age group, where the age group is consistent with the age grouping in the residual correction above. For example, if the sample proportion of 20-25 years old is 5%, then the weight of this group is 20; if the sample proportion of 70-75 years old is 1%, then the weight of this group is 100. This effectively balances the weights of the older / younger groups with smaller sample sizes, preventing the cardiac age prediction model from being biased towards the middle-aged group with larger sample sizes.

[0128] The Adam optimizer adaptively adjusts the learning rate for each parameter data and combines the first and second moments of the gradient, making it suitable for regression tasks involving non-stationary objective functions such as heart age prediction. The Adam optimizer has an initial learning rate of 0.001, a first moment estimation decay rate β1 of 0.9, a second moment estimation decay rate β2 of 0.999, a numerical stability term eps of 1e-8, and an L2 regularization weight decay of 1e-5 to prevent overfitting.

[0129] The ReduceLROnPlateau learning rate decay strategy uses the validation set weighted mean squared error, rather than the training set loss, to ensure that the learning rate is adjusted based on generalization ability. If the validation set loss does not decrease within 10 consecutive training epochs, the current learning rate is multiplied by a decay factor of 0.5, with the learning rate reduced to a minimum of 1e-7, to prevent the heart age prediction model from stagnating due to an excessively low learning rate.

[0130] Step 303: Validate the cardiac age prediction model.

[0131] like Figure 3 As shown, by monitoring the changes in the loss value of the validation set, an early stopping strategy is adopted to prevent overfitting and ensure the generalization ability of the heart age prediction model.

[0132] After each training cycle, the loss value of the heart age prediction model on the validation set is calculated. A patience value is set. Within this value, if the current validation set loss is less than the historical lowest loss value, the model parameters, including network weights, optimizer state, learning rate, etc., are immediately saved to a local file. If the validation set loss value is greater than the historical lowest loss value, the training process is terminated, triggering early stopping. Finally, the model parameters from the training cycle with the lowest validation set loss value are used as the final heart age prediction model, rather than the parameters from the last training cycle.

[0133] In this invention, the patience value is set to 10, which means that if the loss value of the validation set does not fall below the historical low within 10 consecutive training cycles, the training process is terminated and the model parameters are saved as the final heart age prediction model.

[0134] An early stopping strategy is adopted. By monitoring the changes in the validation set loss, training is stopped when the model's generalization ability reaches its optimum, thus avoiding overfitting caused by overtraining and preserving the optimal model parameters.

[0135] The specific operations of step 4 include:

[0136] like Figure 4 As shown, based on the validation set of a healthy sample set, a mapping relationship between the actual age group and the prediction bias is established. The average prediction residual of each age group is calculated as the bias correction coefficient, forming a correction coefficient lookup table. By capturing the nonlinear relationship between age and prediction bias, an age-specific bias correction basis is provided for subsequent test sets / new samples, eliminating the systematic bias of overestimation for younger ages and underestimation for older ages.

[0137] Step 401: Group the age range by age.

[0138] Let the minimum actual age of the validation set samples be age. min The maximum value is age. max age range min ,age max [Divided sequentially into K consecutive age groups G] j The width of each age group is a fixed value w. The j-th age group G... j for:

[0139] G j =[agemin +j×w,age min +(j+1)×w], j=0,1,…,K-1,

[0140] Where K=(age max -age min ) / w; w is the span of each age group, w=3~10, and is preferably 5 in this invention; j is the group index, used to identify different age groups.

[0141] Step 402: Calculate the deviation coefficient for each age group.

[0142] For each sample i in the validation set, the true age is y. A,i The predicted age of the cardiac age prediction model is y. A,pred,i Iterate through all samples in the validation set, and sort the samples according to their corresponding real age y. A,i Classified into the corresponding age group G j In the middle. For each age group G j Calculate all N contained within it j The average predicted residual of each sample is the bias correction coefficient b for that age group. j :

[0143] .

[0144] Finally, the correction coefficient lookup table {G} is obtained. j :b j}, where each age group G j Each corresponds to a specific deviation correction coefficient b. j .

[0145] The correction coefficient lookup table is constructed using age grouping, sample size within the group, and bias correction coefficient as the corresponding relationships. For example, if the age group is 20-25 years old and the sample size within the group is 100, the corresponding bias correction coefficient is 1.2.

[0146] Step 403: Apply the bias correction coefficient to correct the bias of the sample to be corrected, and calculate the corrected age difference.

[0147] The residuals of the heart age prediction data output by the heart age prediction model are corrected to obtain a correction coefficient lookup table. The corrected predicted heart age is obtained according to the correction coefficient lookup table, thus obtaining the corrected age difference value.

[0148] Specifically, for the sample m to be corrected, its actual age is y. B,m The original predicted heart age of the heart age prediction model is y. pred,m Based on the actual age y of the sample B,mDetermine its age group G. j From the correction coefficient lookup table, find G for this age group. j The corresponding deviation coefficient b j The original predicted cardiac age y of the sample m to be corrected is... pred,m Subtract the corresponding deviation correction factor b j The corrected heart age y was obtained. corr,m :

[0149] y corr,m =y pred,m -b j .

[0150] The corrected age difference (AG) is calculated based on the corrected cardiac age and the original predicted cardiac age. corr :

[0151] AG corr =y corr,m -y B,m .

[0152] In this step, the track division module also adopts a time-axis linkage traceability mechanism, including the correction of associated tracks and correction coefficient tracks. All track data are updated in real time, improving the interpretability and accuracy of the correction process.

[0153] The trajectory segmentation module adds a correction tracing time axis to construct correction-related trajectories, including prediction result trajectories, residual trajectories, and age grouping trajectories. The prediction result trajectories mark the original predicted cardiac age and the corrected cardiac age; in this invention, the original predicted cardiac age is marked in black, and the corrected cardiac age in blue. The residual trajectories mark the original residuals and the corrected residuals; in this invention, the original residuals are marked in red, and the corrected residuals in green. The age grouping trajectories divide age groups into ranges of 3 to 10 years, distinguished by different colored blocks.

[0154] The correction coefficient track displays the deviation correction coefficient b for each age group in real time. j The system displays the calculation process of the average predicted residual for all samples within the age group, and allows users to view detailed coefficient calculations by clicking. When abnormal residual samples are manually corrected on the timeline, the underlying system recalculates the deviation correction coefficients for the corresponding age group in real time and updates the corrected cardiac age and age difference AG. corr .

[0155] Cardiovascular risk stratification based on corrected age differences is used for disease identification and mortality risk prediction. For example... Figure 5 , Figure 6 As shown, the yellow bars represent the sample distribution, and the curves represent the comparison between the validation set and the test set of the healthy sample set. Figure 5 Before correction, Figure 6 This indicates the corrected age. Before correction, both the validation and test sets of the healthy sample set showed bias. However, after using the bias coefficient of the validation set to correct the age of the healthy test set, the age bias disappeared. After correction, the average age difference between each age group was close to zero, eliminating the problem of overestimating younger ages and underestimating older ages.

[0156] like Figures 7-9 The image shown is a heatmap of the cardiac age prediction model. Brighter areas indicate a higher concentration of scatter plot values. The horizontal axis represents the actual age, and the vertical axis represents the predicted age from the cardiac age prediction model. Figure 7 It is the training set of the healthy sample set. Figure 8 It is a validation set for the healthy sample set. Figure 9 This is the test set of a healthy sample set. It can be seen that in the three healthy sample sets, the scatter points are all distributed near the curve y=x, indicating that the heart age prediction model has a good predictive effect. The heart age prediction model achieves a MAE of 7.32 on the test set, with a Pearson correlation coefficient of 0.86 and R²=0.74, which is better than the MAE of the existing technology (8.38).

[0157] After correction, the T-scores for both disease and healthy individuals significantly increased, enhancing clinical discriminative power. The T-score represents the significance of the difference between the cardiac age of the disease sample set and the cardiac age of the healthy sample set. The larger the absolute value of the T-score, the more significant the difference; a positive T-score indicates that the average predicted cardiac age of the disease sample set is greater than that of the healthy sample set; a negative T-score indicates that the average predicted cardiac age of the disease sample set is less than that of the healthy sample set. As shown in Table 2, the increase in the corrected T-score for each age group indicates a greater difference between the healthy and disease sample sets, making it easier to distinguish between diseased and healthy patients; after overall correction, the T-score should change from negative to positive.

[0158] Table 2 Comparison of T-scores before and after correction for disease and health test sets.

[0159]

[0160] like Figure 10 As shown, hazard ratio (HR) prediction is performed by grouping patients by different age differences. A risk level (AG) greater than a threshold is considered high risk, and vice versa. Figure 11 , Figure 12The figures show the survival rate curves before and after correction at the optimal threshold, respectively. The horizontal axis represents the age difference, and the vertical axis represents the hazard ratio. A larger hazard ratio indicates a greater difference in risk between different age groups when using the age difference value for age grouping, making it easier to distinguish between high-risk and low-risk individuals. The optimal threshold before correction was 6, with a maximum hazard ratio (HR) of 1.89; while the optimal threshold after correction was 1, with a maximum HR of 3.33. These results demonstrate that the corrected age difference has a stronger correlation with survival rate and is a superior survival predictor biomarker.

[0161] like Figure 13 As shown, the horizontal axis represents actual age, and the vertical axis represents the age difference. The results indicate that residual correction approaches zero more closely than linear correction, exhibits strong generalization, and can stably correct biases across all age groups. In particular, residual correction outperforms linear correction methods in extreme age ranges.

[0162] This invention presents a deep learning-based method for correcting cardiac age prediction bias. First, it corrects the age-related aging (AG) for each age group using a bias correction coefficient, eliminating overestimation in younger individuals and underestimation in older individuals. This achieves a physiological match between cardiac age and chronological age, freeing the AG index from age-related errors and better reflecting the accelerated aging of the heart relative to chronological age, thus restoring its value as a core biomarker for assessing the rate of cardiac aging. Second, it enhances clinical discriminative power, significantly improving the T-score between disease and healthy groups, and making the distinction between diseased and healthy individuals more significant. This allows the AG to more accurately capture pathological cardiac aging caused by disease, providing a more reliable quantitative basis for early clinical disease screening and disease assessment. Third, the corrected AG exhibits superior mortality risk prediction capabilities, with significantly better risk discrimination than before correction. This provides strong support for precise clinical risk stratification and targeted intervention for high-risk individuals, making it a superior survival prediction biomarker. Finally, it demonstrates strong generalization ability, stably correcting biases across all age groups. Especially in extreme age ranges, the residual correction effect is superior to traditional linear correction methods, effectively solving the prediction reliability problem of the model in the marginal regions of data distribution, laying the foundation for cross-population and cross-institutional clinical applications.

[0163] In summary, the corrected AG not only achieves breakthroughs in technical indicators but also deeply aligns with actual clinical needs. It can sensitively identify potentially high-risk individuals and initiate intervention measures in advance; it can also dynamically monitor AG changes to evaluate the effectiveness of intervention treatments, facilitating its widespread application in primary healthcare institutions and health management scenarios, and possesses extremely high clinical translational value.

[0164] The above description illustrates preferred embodiments of the present invention and helps those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative and should not be construed as limiting the specific implementation of the present invention to these embodiments. For those skilled in the art, several simple deductions and modifications can be made without departing from the inventive concept, and all such modifications should be considered within the protection scope of the present invention.

Claims

1. A method for correcting bias in cardiac age prediction based on deep learning, characterized in that, include: Step 1: Obtain 12-lead ECG data and actual age information, filter the ECG data, and construct a healthy sample set and a disease sample set; Step 2: Preprocess the acquired ECG data; Step 3: Build a cardiac age prediction model and train and validate it; Step 4: Perform residual correction on the heart age prediction data output by the heart age prediction model to obtain a correction coefficient lookup table. Obtain the corrected predicted heart age according to the correction coefficient lookup table, and thus obtain the corrected age difference value. Step 2 includes: Step 201: Resample the 12-lead ECG data to 500Hz and crop it to a fixed duration of 10 seconds; Step 202: The amplitude of the 12-lead ECG data signal is normalized and arranged in the order of I, II, III, aVR, aVL, aVF, V1-V6 to form the model input data; The heart age prediction model includes an initial convolutional block, four stacked residual blocks, a global average pooling layer, and a fully connected output layer. It takes actual age as input and heart age as output, uses weighted mean squared error as the loss function, an Adam optimizer, and a ReduceLROnPlateau learning rate decay strategy to train the heart age model. The model monitors the change in loss values ​​on the validation set and employs an early stopping strategy to prevent overfitting. Step 4 includes: Step 401: Group the age range by age; Step 402: Calculate the deviation coefficient for each age group; Step 403: Apply the bias correction coefficient to the sample to be corrected to correct the bias, and calculate the corrected age difference value. Age range min ,age max [Divided sequentially into K consecutive age groups G] j The width of each age group is a fixed value w, and the width of the j-th age group G is... j =[age min +j×w,age min + (j+1)×w], j=0,1,…,K-1, w is the span of each age group, w=3~10, for each sample i in the validation set, the true age is y. A,i The predicted age of the cardiac age prediction model is y. A,pred,i The samples were sorted according to their corresponding real ages y. A,i Classified into the corresponding age group G j In the middle, for each age group G j Calculate all N contained within it j The average predicted residuals of each sample are the bias correction coefficients for that age group. The correction coefficient lookup table {G} is obtained. j :b j }; For the sample m to be corrected, its actual age is y. B,m The original predicted heart age of the heart age prediction model is y. pred,m Determine its age group G. j Find the G value for this age group from the correction coefficient lookup table. j The corresponding deviation coefficient b j The original predicted cardiac age y of the sample m to be corrected pred,m Subtract the corresponding deviation correction factor b j The corrected heart age y was obtained. corr,m =y pred,m -b j ; The corrected age difference (AG) is calculated based on the corrected cardiac age and the original predicted cardiac age. corr =y corr,m -y B,m .

2. The method for correcting cardiac age prediction bias based on deep learning according to claim 1, characterized in that, Step 1 is as follows: Step 101: Obtain multi-source data from multiple databases; Step 102: Select normal samples according to clinical standards and remove noise, artifacts and missing signal samples from the 12-lead ECG data; Step 103: Perform age balancing on the normal samples to construct healthy sample sets and disease sample sets respectively.

3. The method for correcting cardiac age prediction bias based on deep learning according to claim 2, characterized in that, Samples of subjects aged 16 to 85 years were selected for age balancing, and samples were randomly drawn from each age group until the total sample size of 2000 for that age group was reached; if the sample size of a certain age group was insufficient, all samples were included. The healthy sample set was divided into training set, validation set and test set in an 8:1:1 ratio, and the distribution ratio of each age group in the training set, validation set and test set remained consistent. The disease sample sets include LBBB sample sets, RBBB sample sets, HF sample sets, and LVH sample sets.

4. The method for correcting cardiac age prediction bias based on deep learning according to claim 2, characterized in that, When the application layer receives ECG data, it calls the underlying artifact detection request entry module to parse the ECG data, separates the ECG signal and custom configuration parameters, and sends the ECG signal to the dual-dimensional merging processing module and the custom configuration parameters to the artifact priority configuration module. The dual-dimensional merging processing module shares the judgment parameters for similar artifact instances within a continuous time interval of the same lead, and only records the start and end times; for artifact instances in different leads but with matching features, it uses unified detection logic to generate detection clusters and passes them to the hierarchical storage management module; the hierarchical storage management module stores the core lightweight instances in the high-speed cache, and stores the merged detection clusters and custom threshold instances in the secondary storage. During detection, the core instance is first loaded from the cache. If no match is found, the detection cluster / low-frequency instance from the secondary storage is loaded. The matching instance / detection cluster is then sent to the hierarchical voxelization detection module. The artifact priority configuration module combines clinical default rules with user-defined configurations to generate a priority strategy for the current detection and sends it to the graded voxelization detection module. The graded voxelization detection module performs high-precision detection on high-priority artifacts and voxelization detection on low-priority artifacts based on the received artifact instances / detection clusters and priority strategy. At the same time, it adjusts the detection precision according to system resource usage and outputs the artifact detection results to the application layer.

5. The method for correcting cardiac age prediction bias based on deep learning according to claim 1, characterized in that, Arranged in a fixed order of I, II, III, aVR, aVL, aVF, V1-V6, the 12-lead data signals are reorganized to form a (12, 5000) signal matrix.

6. The method for correcting cardiac age prediction bias based on deep learning according to claim 1, characterized in that, Step 3 includes: Step 301: Construct a cardiac age prediction model; Step 302: Train the heart age prediction model; Step 303: Validate the cardiac age prediction model.

Images